m1keil

For now..

Waiting for QuackDB the OSS spin off any second now

huh.. I'm a bit of local LLM noob so I wasn't familiar with mlx_vlm.

I gave it a shot now:

mlx_vlm.generate --model mlx-community/Qwen3.8-27B-4bit --prompt 'give me fizz buzz in rust' --enable-thinking --draft-kind mtp --draft-model mlx-community/Qwen3.8-27B-MTP-4bit --verbose

==========

Prompt: 58 tokens, 90.717 tokens-per-sec Generation: 145 tokens, 36.392 tokens-per-sec Peak memory: 17.419 GB Speculative decoding: 2.79 accepted tokens/round (1.79 accepted drafts/round, 89.4% of drafted, avg draft 2.00) over 52 rounds

Which is very close to ollama, thank you!

I'm not sure if I can get rid of the drafter model, if I understand correctly, the Qwen model already includes a built in draft headers, but just having --draft-kind mtp results in about 17 t/s.

I have a 36gb M3 Max. I tested it across quite a few different options: llama.cpp, oLMX, ollama with different options.

So far ollama managed to be the most performant of them all. I will get 30 to 40 tokes/sec with it when using the -mlx version of Qwen3.8.

Whatever the sauce the ollama folks baked into the mlx + MTP mix is currently working the best out of the box.

Can be easily the price you will pay across Australia, depending on the time of day.

The problem is it's not the engineers that overthink it. The requirement for soc2 usually comes with the first "serious" customer. It is usually a big blocker on some fat contract and now the business makes it your problem for the next 6 months.

So what do you do? You engage and some 3rd party 1800-need-soc2 clowns which will hold your hand and implement all the cookie cutter solutions they know will make auditor happy (oh and btw, they know the auditor personally).

No, that's not a brag but a suggestion to take things into context and not apply soc2 as a cookie cutter solution where a 20k employee enterprise and a 20 people startup must share commonalities.

In a 20 people startup it's very likely that most engineers have access to production anyway and can inject malicious stuff directly, so PRs stop no one really.

It's an old phone getting a second life, the health of the battery is probably not a big concern at this stage.

Heh.. Russia is catching up with Australia!

TIL waydroid...

I don't think that guesstimating that something is an AI slop is very harsh these days. It sure has the signs of one:

- HN User created 66 days ago

- In these 66 days it submitted 4 stories about this project

- The heavy lifting is all done by v86 project (not jslinux as initially suspected)

- The project is a "rebrand/fork" from another project by the same author - "traits.build".

- In the original project, the first commits that mention v86 seem to start at around 23rd of April.

To be clear, I don't really mind people vibing & slopping around. I just would want to have a disclosure about it.

Probably a slop of https://bellard.org/jslinux/

Why this has to be all or nothing?

You can use the drug to loose weight while trying to understand the underlying problem.

Isn't multiphase is Ubuntu only?

Open as in single API layer which allows you to swap the model under it.

A related article from fireworks.ai about running open weights models and why such verifier needs to exists in the first place

https://fireworks.ai/blog/quality-first-with-kimi-k2p5

SOC2 is quite a racket on its own so I'm not surprised to read this industry creates players like this.

I hope that with LLMs, answering security questionnaires will be much less time consuming for companies and less would opt out to get a full blown SOC2 cert. But it will probably play the other way.

Pragmatic and practical. I learned something, thanks.

> Say you have a Bazel project that builds a web application

Ok, wait, why?

They had quite a few release in the last year so it's not dead that's for sure, but unclear how many new customers they are able to sign up. And with IBM in charge, it's also unclear at what moment they will loose interest.

Looks lovely.. I'll definitely will give it a try when time comes.

Nomad is great, but you will still end up with a control plane.

I'm highly skeptical of this claim as well. Going through NATGW with EIP or auto-assigned IP is the exact same cost for the actual traffic.

O3 seems to think this is Swedish Space Corporation.

I hope we will exit this stage of magic spells and incantations sooner rather than later.

Examples were useful to visualise the difference, thanks.

Anyone used both Kagi assistant and perplexity and can share how was the experience?

How is the camera quality on the GrapheneOS phones?

> This is more about guiding the readers, making sure the expectations are crystal clear and that they can follow me throughout an explanation.

Sure, but this holds true for the blog version as well, right?

To be clear, I'm not advocating for The Little Schemer version, and am not arguing that the blog version is the best it can be, but surely we can agree that book padding phenomenon does exist.

By the way, I have read parts of your book over at O'Reilly Learning, and I do think it is a good book. So I'm not trying to take a dump on your work. My criticism is aimed at publishers.

> I learned a ton from my DE. Like, really, a ton. Before that, I had been writing on various blogs for about a decade, but writing online is all about being direct because most people don’t have time. With a book, it’s different. People made a deliberate decision to buy your book. Now, it’s your job to bring them somewhere valuable. And if that takes time (meaning more words), so be it.

I have a hard time with this point. It feels to me like a lot of books have A LOT of unecassery padding all over the place.

The example of taking 28 words and turning it to 120 is pretty good at showing this. The first paragraph is totally pointless - we are reading a book about 100 most common mistakes, obviously this mistake is very common, how did this increased the value?

Then we have another line that explaining what happens in the code, which is totally useless because the code is super trivial.

Then the code, with more explanations on the side as if the previous line was not clear.

And only after that we get to the crux of the issue.

I understand that book publishers feel they need to justify the price of a book by reaching the 300p mark in some or other way, but in my way this only makes the book worse.